There are many catastrophic events occurring in the world. Some of these are due to the use of artificial intelligence in controlling robotic systems whilst others are natural disasters. Catastrophes caused by the use of artificial intelligence in robotics can often be avoided with the successful application of safe artificial intelligence and the damage caused by natural disasters can often be mitigated using the predictive power of machine learning. This thesis explores novel uses of artificial intelligence in reducing the damage caused by catastrophes in robotic control and meteorological prediction. First, a formulation for inverse reinforcement learning (a method of learning by observing an expert) is presented which allows the expert to display varying levels of carefulness depending on how likely they perceive a catastrophe to be. Second, it is shown that misaligned weather datasets can be cleaned using a neural network, without requiring access to a clean dataset for training. Third, it is demonstrated that the prediction of tropical cyclone tracks can be accurate without having to use a generative model and that inference time is much shorter with a non-generative approach.
Catastrophic events in robotics and climate
Hanslope, J. R. P. (Author). 20 Jan 2026
Student thesis: Doctoral Thesis › Doctor of Philosophy (PhD)